I would measure good quality as follows: 1. Completeness: All the core points are covered. 2. Concise: Ignore the unnecessary chatter and retain focus on the points to be noted for the future. 3. Volume of data accepted: Some meeting notes can be very large. Quality here refers to how much data can be accepted without skipping or corrupting the content. 4. Hallucinations: The output should not create new content that was discussed or draw inferences that are not intended. 5. Visual output: The output in terms of diagrammatic flows should be easy to review at a glance and should not be gibberish or too short to understand.

Product: Grammarly 1. What's the real job this solves? Correct spellings, grammatical errors, tone and sentence formation. This helps users focus on the intent rather than on typing. 2. How does the system stay grounded? The system only focuses on spelling, tone, and sentence structure. It does not try to summarise, deduce meaning, or train on the data. It does not aim to teach the language, replace or suggest better words. It only focuses on the text being typed and does not try to correct other existing text. 3. What's the context the model receives? The context is the text being typed actively or the input space. All other existing text is ignored. By input space, I mean a Word document being edited, an input box in a chat (not the whole conversation), or an email body, not the rest of the email structure. 4. What are the failure modes? It only provides suggestions, putting the onus on the user to accept or reject them. It allows the user to alter the tone to provide better answers. It does not hallucinate, but with limited context, the tone it detects and the suggestions it provides are sometimes incorrect. 5. What trade-offs did the team make? They selected accuracy over speed, as the suggestions are almost always correct for the given context They chose consistency over flexibility as there is only one answer. They chose safety over usability since it does not seem to detect some text. 6. What's the role of the UI? UI is simple. A blow-up bubble near the input with the number of suggestions, non-interfering with text input. The suggestions appear over the highlighted text, so the user can compare the old and new text. The user can select the text by simply clicking on the suggestion- intuitive design, as the user does not have to search for an accept button. A simple dismiss button is all the user needs to reject the suggestion. It also indicates whether it is a spelling, punctuation, tense, or other issue, informing the user about the language semantics used to correct the text.

Product: A trusted support system and confidant for new mothers. 1. Data Readiness : 4/5 Some data, but it's patchy. It uses LLMs' existing knowledge to build upon. There is a lot of data available on maternal health and baby care, published by the National Boards for Maternal Health, with the latest, most relevant region-specific information. LLMs can read data from wearable devices to infer objective physiological signals that can be used to understand the mother's mood and health. This data can be stored to analyse behavioral patterns and accordingly provide assistance as needed. Users can provide local maternal insights to the app, which are vetted by a human before being used as training data. Local maternal insights are undocumented and will be collected as the app is used. There is no training data based on conversational patterns. Using conversions from one user to train the model is not suitable since each user is unique. But it can be reviewed to check for hallucinations and for safety. 2. Output type : 5/5 Subjective/judgment-based. The responses are totally subjective and depend upon multiple parameters such as context, mothers mood, time of the day, and the users' behavioral patterns. 3. Error Tolerance : 4/5 The output is conversational. It is up to the mother to use it or not. To capture hallucinations, the user must reject such replies. I will not rate it a 5/5 because it should not misguide mothers. And hence, guardrails will need to be built around the conversations. Total score: 13/15. Its definetely ideal for AI. Implementing safety rules is something Im not very sure about at this stage.

@Akshun Gulati This is a common problem and good to solve. As correctly identified, you will need to use embeddings to identify all the different terms used for the same thing. One of the problems here is interchangeable terms. Same terms are used in different context by different invoices. You can expand the solution to include different countries and languages too. Secondly, the structure and layout may also vary greatly which means you will need a lot of data for training.

The problem: An app that helps new mothers track their health and their baby's reactions. 1. LLM - The solution needs to learn from data from various reliable sites on maternal and child care. It needs to learn about patterns in maternal health and baby care to provide timely advice. The conversational model must identify the mother's mood and accordingly set the tone of the conversation to calm them and provide support. 2. RAG - As the user continues to use the solution, their medical history, milestone dates, timeline events, and behavioural patterns will continue to be added to the database. This would provide a base for future investigations when needed. Such as when a rash appeared, probably the time when a new pair of shoes was introduced. A database of emergency contact numbers, hospitals and other help providers is required in case of an emergency. 3. Embeddings - No embeddings are required here since it is a conversational model. 4. Other - I would need to train the model to read, store and analyse data from wearables. Also, train the model on various traditional maternal remedies that may benefit the mother, but the data for this is extensive and may be mixed with superstition, which needs to be cleaned up.

PROBLEM I’M THINKING ABOUT An AI personal caretaker that supports new mothers and babies by providing timely, contextual physical and mental health guidance, tracking changes, and proactively flagging potential concerns. Q1 – WHAT PROBLEM? Despite abundant information on maternal and newborn care, parents face: - Fragmented and generic information that isn’t personalized. - Lack of mental health support for new mothers and confused partners - Cognitive overload and sleep deprivation, making it hard to identify why a baby is uncomfortable or crying. Subtle environmental or routine changes often go unnoticed, leading to anxiety and trial-and-error parenting. Q2 – SHOULD WE USE AI? Yes. New parents need 24×7, context-aware support. AI can act as an always-available baby and maternal care companion that: - Answers questions in real time - Tracks routines, changes, and patterns over time - Reassures parents about what is normal and flags deviations - Adapts advice based on cultural, regional, and family context This level of personalisation and continuity is not possible with rule-based systems. Q3 – DO WE HAVE THE DATA? Yes. Data is created naturally through usage: - Conversations with the system - Baby routines (sleep, feed, mood) - Mother’s self-reported physical and mental health - Voluntary inputs such as photos or notes Foundational medical and child-development knowledge already exists to support the base system. Q4 – TRANSLATE TO A MODEL All users start with a common base model.Each family’s data is used only to personalize their experience and is not shared across users. The model: - Maintains conversational context - Tracks changes over time - Predicts likely causes of discomfort or stress - Suggests safe, non-diagnostic remedies and clearly escalates when needed Admins control base knowledge and safety boundaries. Q5 – USER EXPERIENCE - A shared dashboard showing baby and mother health indicators - Private conversations for each family member - Proactive tips on what’s happening now and what to expect next - Milestone tracking with suggestions for activities and toys. - Reduced engagement and user feedback signals loss of relevance, prompting the system to adapt its approach.

@Sid Arora Thank you for the inputs. The wearables idea is great. Most tasks are repetitive such as feeding, diaper change, baby massage, putting them to sleep. If the wearable can detect patterns in hand movement, then the mom does not have to log the baby's routine! Moreover, the wearable can detect the mother's wake and sleep cycles and fatigue providing helpful tips to manage her health and when needed connect to her doctor for timely advice. Yes, I also share the concern about medical advice given by AI and overly relying on it. My thoughts are it should only be a recommendation.